The evidence base on AI in K-12
A · Strong evidence A Stanford SCALE evidence review co-authored by education economist Susanna Loeb that screens 818 papers down to 20 causal studies (mostly RCTs) using What Works Clearinghouse standards.
Summary based on the full text·Original: about 50 minutes to read
In short
A review by Lily Fesler, JP Martinez Claeys, Chris Agnew and Susanna Loeb at the AI Hub for Education of Stanford's SCALE Initiative (2026). It examines the 818 papers in the Hub's repository as of October 2025 and finds only 20 provide strong causal evidence on how AI affects students and teachers, none of them student-facing studies in US K-12 schools. The evidence suggests AI improves student performance while students have the tool, with mixed or negative effects once it is removed, and that AI can save teachers time and raise the quality of less experienced tutors.
Key numbers
- 818 / 20papers in the AI Hub research repository (October 2025) vs papers with strong causal evidence
- about 30%less time on lesson preparation for teachers given ChatGPT and a guide, with no detectable drop in quality
- 4 to 9 ppgain in topic mastery from Tutor CoPilot overall, rising to 9 points for students of lower-rated tutors (RCT, 900 tutors)
As reported by the source.
English·PDF·opens scale.stanford.edu
What it reports
- Repository growth: 28 qualifying papers in January 2023, over 800 by October 2025; the count doubled between January and September 2025.
- Study designs in the repository: descriptive 46%, technical/computational 46%, RCTs 8%, quasi-experimental 5%. 64% are in postsecondary settings.
- Fourteen causal student studies: seven university experiments (three US, four Germany) and seven with high school students (Turkey, UK, Germany, Belgium and Spain, Brazil).
- Students using AI score higher during practice, but in Turkey students who practised with a general-purpose chatbot did worse on the unaided exam; a tutoring-specific chatbot produced no difference.
- A German experiment found students researching with a general chatbot showed lower-quality reasoning than those using a search engine.
- Teachers given ChatGPT and a guide spent about 30% less time (about 25 minutes per week) on lesson preparation with no detectable quality difference in blind ratings; AI use fell from 39% to 29% of lessons while savings persisted.
- Tutor CoPilot (RCT, 900 tutors): mastery up 4 percentage points, 7 points for less experienced tutors, 9 points for lower-rated tutors. Google's LearnLM raised performance on later topics by 5.5 points.
How they got it
The authors took the October 2025 snapshot of the repository, mostly arXiv preprints matching "education" and "AI". An LLM pre-classified papers for causal strength using What Works Clearinghouse criteria; at least two human researchers then reviewed each candidate. Findings are synthesised through learning science principles such as cognitive load and desirable difficulties.
Limits
- Most underlying papers are preprints; keyword search may miss journal work and favours recent papers.
- Studies are short, often one-off sessions, and mostly outside US schools, so transfer to classrooms is uncertain.
- Equity and student wellbeing effects are largely unstudied.
Why it matters for companies
The distinction between tool-assisted performance and durable learning applies directly to corporate training and onboarding: staff who learn with a general chatbot may look productive while building less skill. Tools that scaffold reasoning rather than hand over answers performed better. The tutor results suggest AI coaching helps the least experienced staff most, a pattern seen in customer-service studies too. Companies measuring AI learning tools should test people without the tool, not only with it.
Summary based on the full text. Read the original for the full argument and data.Open PDF on scale.stanford.edu